I have developed a Pix2Pix GAN that generates maps from satellite images. My problem is that the maps do not look perfect, and sometimes even rather awful. Can someone help me? My hyperparameters are:
BUFFER_SIZE = 1096
BATCH_SIZE = 1
IMG_WIDTH = 256
IMG_HEIGHT = 256
lr = 0,0002
beta1 = 0.5
beta2 = 0.999
lambda = 100
The Discriminator Accurancy is : 83 %
Welcome to the TensorFlow Forum!
There are also some other factors involved to get refined results in GAN model like model architecture and the number of epochs/training steps for the model training which will affect the result. Please have a look at this TF pipx2pix GAN model for the reference. Here the model training done for more than 4k steps and 200 epochs as well as used restoring the checkpoints to have better results. Thank you.